NTT123/vietTTS

Combines a three-stage neural architecture—duration prediction, acoustic feature generation, and HiFiGAN vocoding—to synthesize Vietnamese speech from text. Trained on the denoised InfoRe dataset with forced alignment via Montreal Forced Aligner, it supports model finetuning on ground-truth mel-spectrograms and experimental multi-speaker synthesis on a separate branch. Implemented in JAX/Haiku with PyTorch vocoder conversion, enabling both offline synthesis and integration into production pipelines via pretrained model inference.

255 stars. No commits in the last 6 months.